Attention Based Hybrid Deep Learning models for Multi class Amharic News Categorization with Explainable AI

Efficient and adaptable text classification systems are required due to the increasing expansion of Amharic electronic documents in order to manage and extract important insights from large amounts of data. Traditional natural language processing (NLP) methods are severely hampered by the Amharic language's intricate morphology and sparse annotated corpora. By combining the advantages of Bidirectional GRU and Bidirectional LSTM networks for capturing sequential dependencies and Convolutional Neural Networks (CNN) for local feature extraction, this study proposes a novel attention-based hybrid deep learning model for multi-class Amharic news classification. The model's discriminative power across several categories is improved by incorporating a self-attention mechanism that dynamically emphasizes contextually significant terms. We employ Explainable AI (XAI) methods, like Local Interpretable Model-Agnostic Explanations (LIME), which offer human-interpretable explanations for predictions, to minimize the hidden nature of deep learning decisions and enhance trust in AI systems. A carefully selected multi-class Amharic news dataset that encompasses a variety of topics, including agriculture, culture and tourism, education, the economy, the environment, foreign affairs, health, politics, science and technology, and sports, is used to evaluate the model. The experimental results show that CNN + BiLSTM with Self-Attention scores 96.7, 96.8, 96.8 and 97 for precision, recall, F1-score, and accuracy respectively. This study introduces to the larger objective of transparent and reliable AI in low-resource language contexts and pushes beyond the limits of Amharic automatic language processing.